This paper proposes a dual-stream collaborative recommendation approach based on BERT4Rec and user session enhancement to address the disconnection between long-term interests and short-term session dynamics and to reduce response lag in e-commerce recommendation systems. This paper innovatively constructs a synergistic mechanism combining time-aware enhancement and gated fusion. For short-term session enhancement, a time-aware attention mechanism is introduced. By combining a temporal decay factor with content relevance, composite attention weights are dynamically calculated to precisely enhance recent high-intent actions. A gated fusion unit is further designed to adaptively integrate long- and short-term representations, achieving a personalized and dynamic balance of interests. This paper precomputes and caches computationally expensive long-term interest vectors offline, performing real-time inference only on the current short-term session path. Approximate nearest neighbor retrieval is then combined to accelerate candidate generation, balancing model complexity with the low latency requirements of online services. Experimental results demonstrate that proposed method achieves excellent recommendation accuracy (HR@10 and F1@10 of 0.438 ± 0.008 and 0.402 ± 0.009, respectively) and ranking quality (average NDCG@10 (Normalized Discounted Cumulative Gain at 10) of 0.551 ± 0.004 at Epoch 50). Recall@5 reaches 0.401 ± 0.013 among infrequent users, effectively enhancing the accuracy and responsiveness of recommendations.
Fu et al. (Thu,) studied this question.